Fix a disk handling bug triggered by failover

This leaves an instance's disks configured for the primary node as after
disk activation we want to start the instance anyway. As such,
_GatherBlockDevs in backend.py will need the disks configured for the
primary.

Reviewed-by: imsnah